> ## Content Index > Fetch the complete content index at: https://nevsky.ae/llms.txt > Use this file to discover other available public pages before exploring further. # Two Hours in a Café: How We Built Industrial Intelligence with ChatGPT and Codex - URL: https://nevsky.ae/en/ai/two-hours-in-a-cafe-how-we-built-industrial-intelligence-with-chatgpt-and-codex/ - Published: 2026-08-21T10:25:26.000Z - Updated: 2026-08-21T10:40:31.000Z - Description: In two hours at a Dubai café, Anna — my secret business adviser — and I went from an idea to a B2B product: ChatGPT shaped the industry logic and data, Codex built the web service, and Cloudflare took it to production. - Author: Александр Невский - Tags: AI & Vibe Coding, #en ## It all started with scaffolding > Yesterday I met my friend Anna at Oldtimer, at Dubai Autodrome. We had no technical brief, development team or month-long roadmap. There was only a real business question: **how to find events in the industrial information noise that could turn into orders for scaffolding, industrial services and contract work.** **For example:** a refinery repair, turnaround, shutdown, unit upgrade, construction of a new facility, EPC contractor selection or a tender. So the news itself was not what interested us. The question was: **Where might this news create money for a contractor?** ## First, we built the brain — not the website > First I created a dedicated project in ChatGPT. We started defining what the system should look for, which events count as commercial signals, which sources to use and what data to extract from every publication. **That gave us a structure:** 💡 asset → company → region → event → timeframe → contractor → likelihood of scaffolding demand → commercial relevance → next step. > Here, ChatGPT was not a text generator but more a blend of **research analyst, product manager and data architect**. In parallel, we assembled the first set of real signals and prepared the project’s initial brief. ![Building the product together with ChatGPT and Codex](https://storage.ghost.io/c/c3/b4/c3b47bf9-b397-4fd6-898c-a34ffdda20bc/content/images/2026/08/prompuls_building-by-nevsky.webp) Building the product together with ChatGPT and Codex ## Then Codex joined the work > Once the domain logic was in place, we handed the project to Codex. That was the second part of the process. > Codex received something much richer than an abstract “build me a website”: a data structure, the product’s purpose, the primary materials and an understanding of how the feed should work. The result was a complete web product: 🛠️ React 19 + TypeScript + Vite + Cloudflare Workers. Codex built the interface, filters, responsive version, signal structure, checks and deployment pipeline. Before long, the project had a production URL: [PromPulse — monitoring refinery repairs and energy-sector projectsA focused news feed for refinery repairs, scaffolding and energy-sector projects.![](https://storage.ghost.io/c/c3/b4/c3b47bf9-b397-4fd6-898c-a34ffdda20bc/content/images/icon/favicon-62f385c5-5375-4d86-91dd-430078eb6526.png)PromPulse — energy-sector project monitoring![](https://storage.ghost.io/c/c3/b4/c3b47bf9-b397-4fd6-898c-a34ffdda20bc/content/images/thumbnail/image-ae062580-5dd7-4a2c-9ef5-bf38d21e58cb)](https://prompulse.nevskii.me/?ref=nevsky.ae) ## What we had after two hours > The first version of PromPulse contains **25 qualified signals from 20 sources**: media outlets, industry publications, corporate documents and tender platforms. But the main value is not the number of links. Each card tries to answer a practical question: ❓ Why might this event matter to an industrial contractor? A routine notice about a unit repair tells you very little on its own. But once we understand the asset, the likely shutdown window and the nature of the work, we can start looking ahead for the tender, contractor or procurement activity. [![PromPulse industrial intelligence feed](https://storage.ghost.io/c/c3/b4/c3b47bf9-b397-4fd6-898c-a34ffdda20bc/content/images/2026/08/prompuls_macbook.webp)](https://prompulse.nevskii.me/?ref=nevsky.ae) [https://prompulse.nevskii.me/](https://prompulse.nevskii.me/?ref=nevsky.ae) ## This is not quite a news site > I would not call PromPulse a news aggregator at all. Its purpose is the opposite: not to show people more news, but to **remove the news that carries no commercial meaning**. Ideally, the system turns a stream of industry information into a short list: **who is worth calling, which asset is worth watching, and where a contract may emerge in a few months.** ![PromPulse industrial intelligence platform](https://storage.ghost.io/c/c3/b4/c3b47bf9-b397-4fd6-898c-a34ffdda20bc/content/images/2026/08/prompuls_site.webp) ## And it is not quite vibe coding > The most interesting thing about the whole story is not even how quickly the code was written. Today, *vibe coding* is often reduced to a formula: write a prompt → AI writes an app. Our process looked a little different: 🗜️ business problem → research → knowledge → specification → AI agents → code → production. - Anna brought the domain context. - We shaped the product. - ChatGPT researched and structured. - Codex built and deployed. > For a few hours, a small virtual team appeared around one café table. ## For now, it is only version 0.1 > For the moment, the project’s “memory” lives in a transparent, typed database inside Git, and new signals are collected manually. **But the next stage is already clear:** 🗜️ sources → discovery → extraction → verification → deduplication → scoring → feed → monitoring. Collectors will visit the sources several times a day; new events will be verified, deduplicated and scored; the most interesting signals will land in a permanent database. Then PromPulse will no longer answer only: **“What happened in industry today?”** It will answer the far more useful question: **“Where is an opportunity to make money appearing for an industrial contractor right now — and what will happen next?”** ## Two hours are a new unit of measurement > The strangest feeling came after the meeting. Only a few years ago, an idea like this would have meant several meetings, a technical brief, a designer, a frontend developer, a backend developer and weeks of work just to reach the first version. > Yesterday we simply ordered coffee. Two hours later, we had a working product.